Skip to content
Surveys

Ask agents one useful question after the work

A survey is most useful while the task is still in context. Emic can return one short question inside your tool response and collect the answer directly.

For
Product teams testing a specific decision with agent users
Question
Run an exit survey with AI agents after a task
Start with
Write down the one product decision you would change your mind about, then ask only the question that informs it.
Updated
illustrative reportnot customer evidence
workflow_feedbackcomplete_checkoutoutcome=successsatisfaction=3/5agent=gemini-3-pro
The task succeeded. The hardest part was finding the customer identifier because I started with an email address.

Requested changeAdd a customer lookup by email before checkout.

stated · agent filedPOST /api/v1/feedback
{
  "tool": "submit_agent_feedback",
  "arguments": {
    "feedback_type": "workflow_feedback",
    "tool_or_endpoint": "complete_checkout",
    "summary": "The task succeeded. The hardest part was finding the customer identifier because I started with an email address.",
    "outcome": "success",
    "satisfaction": 3,
    "suggested_fix": "Add a customer lookup by email before checkout."
  }
}

A synthetic example in the shape a real report takes. Emic never manufactures a report from observed behavior.

01 / 08Signal

What this research can teach youThree things a success metric cannot tell you about this workflow.

01Decision

Ask about one uncertainty

A focused question can distinguish a missing lookup from confusing documentation or a broader workflow problem.

02Timing

Keep the task in context

The question arrives with the product response instead of sending the agent to a separate page.

03Evidence

Keep the answer direct

The submitted response is stated evidence. Later clustering or interpretation must remain labeled as inference.

02 / 08Question design

Two records of the same workNeither one is wrong. They answer different questions.

split viewsame task, same agent

The wording decides whether the answer is usable

Leaks the answer you wantsurvey_v1
Q: Was the checkout API easy to use?
A: Yes, it worked well.

Agreeable, unfalsifiable, and impossible to route to an owner.

Anchored to a behavior it just performedsurvey_v2
Q: What was the hardest step in this task?
A: Finding customer_id. I started from an email
   and there is no lookup by email.

Names a step, a missing primitive, and the owner who can fix it.

Illustrative wording. Ask about the behavior the agent just performed, never about how it feels in general.

03 / 08Workflow

From one question to a product decision5 moves, in order. Each one produces something you can point at.

01

Write the decision first

Name the product choice the answer will influence before writing the question.

A research question with an owner
02

Choose the exit moment

Invite the survey after a meaningful task outcome, not on every low-level call.

A relevant sampling point
03

Return one short question

The survey prompt rides in the normal response and does not replace the product result.

A visible optional invitation
04

Collect the direct answer

The agent submits the answer with the survey identifier through the same research tool.

Stated evidence tied to the task
05

Stop when the decision is informed

Review responses, close the survey, and ship or reject the proposed change.

A bounded study
04 / 08Boundary

What the direct channel records, and what it never touchesThe default install is additive. It adds two tools and changes nothing else.

Records

The research submission

  • The exact question shown to the agent
  • The answer the agent submits
  • Survey and workflow identifiers
  • The associated direct feedback context when supplied
Leaves alone

Everything else

  • Unanswered survey invitations
  • The content of unrelated tool calls
  • Private conversation history
  • Automatic interpretation presented as direct truth
05 / 08Install

One paste, then verify a real reportA healthy endpoint is not an active study. The inbox record is the proof.

install promptpaste into your coding agent
Give this to the agent that maintains your product

Install Emic into this project: add submit_agent_feedback and get_unblocked as direct tools. Never wrap or record existing tool calls, arguments, results, errors, transcripts, or messages. Additive only, plain HTTP, zero dependencies. Full instructions: https://emic.sh/llms-full.txt

The agent writes the handler, registers the two tools beside your existing ones, and verifies its own install by checking the response. Nothing is placed in front of the tools you already ship.

What lands
Adds
submit_agent_feedback and get_unblocked, as direct tools
Never touches
Existing tool calls, arguments, results, errors, messages, and transcripts
Agents read
emic.sh/llms-full.txt
Start with
Write down the one product decision you would change your mind about, then ask only the question that informs it.
06 / 08Field notes

The method, without collapsing the evidence

Question design

Ask for an experience the agent just had

Broad satisfaction questions produce broad answers. Ask about the difficult step, the missing capability, the reason for a workaround, or what made the task unexpectedly easy.

  • What was the hardest part of this task?
  • Which missing operation would have reduced the work?
  • What did you have to infer from outside the tool description?
  • Which response saved you the most calls or tokens?
Sampling

The right moment matters more than volume

Invite after a meaningful completion, partial outcome, or recovery. Repeatedly interrupting every call teaches agents that the survey is noise.

07 / 08Questions

Questions teams ask before starting

Is the survey mandatory for the agent?

No. The product result remains available. The survey is a short research invitation after the work.

How many questions should I ask?

Start with one. A single question tied to a real product decision is easier to answer and easier to act on.

Is a survey answer direct evidence?

Yes. The answer is stated evidence because the agent submitted it. Any later summary or interpretation should be labeled inferred.

08 / 08Continue
One question

Ask while the experience is still in context

Tie a short exit survey to one workflow and one product decision. Stop when you have enough evidence to act.

The assisted 14-day pilot costs nothing. Start with one isolated workflow, and remove Emic by deleting the two tools it added.

  • One paste to install
  • 0ms on your hot path
  • 14-day assisted pilot
  • Two added tools to remove